CONTEXTUAL QUERY USING BELL TESTS JOAO BARROS, ZENO TOFFANO, YOUSSEF MEGUEBLI AND BICH-LIÊN DOAN SUPELEC (ÉCOLE SUPÉRIEURE D’ÉLECTRICITÉ) FRANCE Quantum Interaction 2013 Leicester July 25-27, 2013 QUANTUM INTERACTION RESEARCH AT SUPELEC Research activity initiated in 2011 under the impulse of Bich-Liên DOAN (Computer Science: Information Retrieval, Semantic Web), Dep. Of Computer Science. Zeno TOFFANO (Physicist, lectures Quantum Mechanics. Research: Solid State Physics, NMR, Lasers, Fiber Optics Telécommunications), Dep. Of Telecommunications. 2 PhDs Joao BARROS (MsC: Theoretical Physics) at the heart of this research (funding from « Fondation SUPELEC »). Youssef MEGUEBLI (MsC: Computer Science) : Opinion based Information retrieval. We undertook some preliminary investigations in the form of tests. (arXiv:1207.4328) We emphasize on the experimental « Quantum-like » approach. PRELIMINARY INVESTIGATIONS : POLL TESTS ON POLYSEMY IN FOREIGN LANGUAGE: The test assesses correlations in a foreign language (Chinese here). It aims to show the role of polysemy of words. The question was to quantify the correlation with different meanings of the proposed words. 4 people were interviewed (all Chinese) to give their opinion scores. POLL TEST chinese word 笔记本 性 生 清 出入 4 persons polysemy Scores laptop computer 9964 paper notebook 4667 sex 9956 character 2665 life 5585 to be born 9967 Qing dynasty 8648 fair end honest 2563 go in and out 8666 the failure to agree 4696 PRELIMINARY INVESTIGATIONS : POLL TESTS ON HETEROGENEOUS MEDIA The test proposes nine musical excerpts. The question is to rate from 0 to 10 whether these excerpts fall under the category "rock" or "blues ". 4 persons were interviewed. POLL TEST “Music Excerpts belonging to” “Blues” “Rock” + “Blues” Interference 5.5 4.5 10 = 8 2.5 10.5 over 4.5 5.75 10.25 over « Folsom prison blues » 4 4 8 under « The wind cries Mary » 4.75 6.25 11 over « Don't let me down » 8.25 3.25 11 over « Tenth avenue freeze out » 3.5 3.75 7.25 under 2 8 10 = 7.5 3.25 10.75 over « Dazed and confused » « Susie Q » « That's all right » « Since I've been loving you » « I heard it through the grapevine» “Rock” The sum of the results for both categories is very rarely equal to 10 indicating that the chosen categories are certainly not mutually exclusive. Interference effects between concepts of different media (over-extension/under-extension). PRELIMINARY INVESTIGATIONS WORDS BELONGING TO 2 CATEGORIES : CORRELATION Tests on words belonging to two categories Fruit, Vegetable or to both (questions are independent) We define a « Bell-like » correlation parameter 𝑡𝑒𝑠𝑡 𝑆𝐵𝑒𝑙𝑙 = 𝜇 𝐹 𝜇 𝑉 +𝜇 𝐹 1−𝜇 𝑉 + 1 − 𝜇 𝐹 𝑜𝑟 𝑉 − 1−𝜇 𝐹 𝜇 𝑉 In this analysis we observed no violation of the Bell Inequality (<2) POLL TEST “Word belonging to” max: 1 max: 1 max: 1 max: 4 µ(Fruit) µ(Veg.) µ(F or V) SBell garlic .16 .52 .33 0.39 almond .68 .07 .83 0.82 beet .2 .4 .53 0.35 broccoli .04 .92 1 0.93 mushroom .06 .75 .6 0.36 cauliflower .06 .92 .93 0.86 cucumber .18 .72 .83 0.61 gherkin .26 .47 .8 0.41 spinach .04 .95 .8 0.75 bean .1 .97 .87 0.84 coco nut .96 .02 .6 1.36 olive .6 .4 .67 0.77 parsley .14 .35 .47 0.37 pepper .08 .05 0 1.03 potato .16 .45 .17 0.62 apple .94 .03 1 0.94 ginger .1 .22 .26 0.63 grapes 1 .02 1 1 tomato .8 .6 1 0.92 PRELIMINARY INVESTIGATIONS FIRST APPROACH : « HEURISTIC QUANTUM-LIKE HAL MODEL » Our goal here is to classify the texts according to a context-related search criteria using the HAL algorithm (Hyperspace Analog to Language). We create a symmetrical HAL matrix (more discussions hereafter). A query has been undertaken: "A AND B". Texts referring to the word Tomato (can be considered either as Fruit or as Vegetable). Three documents were selected from the Internet using related keywords (for complete discussion see arXiv:1207.4328) The score p given by the algorithm corresponds to the probability for this text to be in first position in our request. The average value of the test is defined as E = 2p1 (E from 1 to +1). A Bell parameter (CHSH type) is defined and calculated as follows 𝐻𝐴𝐿 𝑆𝐵𝑒𝑙𝑙 = 𝐸 𝑝𝑇𝐹 − 𝐸 𝑝𝑇𝑉 We observe Bell inequality violation in document 2 𝐻𝐴𝐿 𝑆𝐵𝑒𝑙𝑙 >2 + 𝐸 𝑝𝑃𝐹 − 𝐸 𝑝𝑃𝑉 Document n° 1 2 3 Tomato AND Fruit 0.788 0.581 0.373 score pTF Tomato AND Vegetable 0.349 0.469 0.213 score pTV Plant AND Fruit 0.651 0 0.385 score pPF Plant AND Vegetable 0.315 0 0.223 score pPV 0.947 2.23 1.105 Bell param. S Long story: The field of Bell inequality violations (Bell 1964) and entanglement has fascinated many scientists throughout the last decades. An interesting historical narrative is in “How the Hippies Saved Physics” by David Kaiser, Ed. W. W. Norton (Physics World 2012 Book of the Year ). Much debate classical and non-classical behaviour entanglement local and non-local, contextual and non-contextual more than Quantum, non-local boxes… Experiments demonstrating Bell inequality violation 1969 Clauser: first experiment 1982 A. Aspect (Orsay France) on polarized photons: definitive proof Entanglement with Spins (NMR, Rydberg atoms…) Towards the realization of a Quantum Computer A new field: Quantum Information Entanglement is at the heart of this field because it is seen as a potential “resource” for computing (lower complexity) and coding (secure cryptography) BELL INEQUALITIES THE BELL CHSH INEQUALITY CASES The CHSH (Clauser, Horne, Shimony, Holt)-Bell parameter SBell form is proposed for tests with two binary outcomes, +1 or 1, adapted to query answers (YES/NO), can be defined as follows: 𝑆𝐶𝐻𝑆𝐻−𝐵𝑒𝑙𝑙 = 𝐸 𝐴, 𝐵 − 𝐸 𝐴, 𝐶 + 𝐸 𝐵, 𝐷 + 𝐸(𝐶, 𝐷) where 𝐴, 𝐵, 𝐶 and 𝐷 are tests and 𝐸 𝑋, 𝑌 stands for the expectation value of the outcome of mutual tests 𝑋 and 𝑌. 𝑆𝐵𝑒𝑙𝑙 can never exceed 4. Classical, local, separable: 𝑆𝐵𝑒𝑙𝑙 lies between 0 and 2. We could write 𝐸 𝑋, 𝑌 = 𝐸 𝑋 𝐸 𝑌 . Quantum: The case 2 ≤ 𝑆𝐵𝑒𝑙𝑙 ≤ 2 2 achieved with bipartite quantum entangled states. 𝑆𝐵𝑒𝑙𝑙 = 2 2 is called the Tsirelson’s bound and is a limit for Quantum systems. No-signalling. The case between 2 2 and 4 is called the “nosignalling” region. The maximum value 𝑆𝐵𝑒𝑙𝑙 = 4 can be attained with logical probabilistic constructions often named non-local PR boxes. HAL AND QI RESEARCH We investigate the relationships between words within a document; these relationships can be formed by creating a “semantic space” using the Hyperspace Analogue Language (HAL) introduced by Lund and Burgess (1996). The HAL algorithm does not require any explicit human a-priori judgment. In the procedure a HAL lexical co-occurence matrix is built with a "window," representing a span of words passed over the corpus being analyzed. Operationally, two words are considered as co-occurring when they appear in the same floating window. The size of this window is a few hits left and right of the word in question. Similar approach: LSA (Latent Semantic Analysis) also builds matrices in semantic space. Darányi, Wittek, Physical analogy between semantic space of HAL and Quantum Theory, where at each word can be associated a given energy (in analogy with spectral emission lines in atoms corresponding to transition energies) Bruza HAL used for analogies with Quantum Theory for activating associations of concepts. THE HAL MATRIX SEMANTIC SPACE The matrix is built with a "window" representing a span of words passed over the corpus being analyzed. The width of this window can be varied. Words within the window are recorded as co-occurring with weight inversely proportional to the number of other words separating them within the window (word distance measure). The information contained in a line is the sum of co-occurrences for words appearing before the word, the information contained in a column represents the sum of cooccurrence for the words appearing after the word. We used a symmetric real positive matrix obtained by the sum of the HAL matrix and its transpose (equivalent to run HAL backwards). All words are considered and simple plurals are treated as singular words. Lower and upper case letters are not distinguished. Words having the same origin are treated differently (for example “battle” and “battling” are distinct). DOCUMENT « ORANGE » CONSTRUCTION OF THE HAL MATRIX Symmetric matrix sum of two HAL matrices (forward and backward). Repeated words contribute to strengthen the associated vector (see “orange” and “the” in the example below). The rows and columns of the symmetric co-occurrence matrix constitute vectors in a highdimensional space. The dimensionality of the space is determined by the number of columns in the matrix (context vectors). TEXT example with a window spanning on 3 words (l = 3) "THE COLOUR ORANGE TAKES ITS NAME FROM THE ORANGE FRUIT" Matrix (l=3) M+M^T THE THE COLOUR ORANGE TAKES ITS NAME FROM FRUIT 16 3 5 1 1 2 3 2 COLOUR 3 8 3 2 1 0 0 0 ORANGE 5 3 16 3 2 2 2 3 TAKES 1 2 3 8 3 2 1 0 ITS 1 1 2 3 8 3 2 0 NAME 2 0 2 2 3 8 3 0 FROM 3 0 2 1 2 3 8 1 FRUIT 2 0 3 0 0 0 1 8 QUANTUM MODEL FOR HAL : VECTOR DEFINITION We attribute to each document an associated vector. The vector state of the document is the linear sum of all the word vectors |𝒘𝒊 it contains. Each word vector state is extracted from the lines of the symmetric HAL matrix. 𝑁 |Ψ = |𝑤𝑖 𝑖 We are interested in analyzing how two words are connected within a document, namely word 𝑨 and word 𝑩. The two associated word vectors |𝑤𝐴 and |𝑤𝐵 define a plane on the semantic space. We will consider the projection of the document vector state |Ψ on the plane spanned by |𝑤𝐴 and |𝑤𝐵 . This resulting normalized state vector, represents the reduced document state vector |𝝍 . QUANTUM MODEL FOR HAL : VECTOR ORTHONORMALIZATION To obtain |𝜓 we take the vectors |𝑤𝐴 and |𝑤𝐵 and normalize them obtaining two new vectors: |𝑢𝐴 and |𝑢𝐵 forming a non-orthogonal basis |𝒖𝑨 , |𝒖𝑩 (in general). We apply the Gram-Schmidt orthogonalization process to the basis {|𝑢𝐴 , |𝑢𝐵 }, and we obtain two new orthonormalized basis that describe the plane formed by the original vectors |𝑤𝐴 and |𝑤𝐵 : the basis {|𝒖𝑨 , |𝒖𝑨⊥ } and {|𝒖𝑩 , |𝒖𝑩⊥ }. By projecting the vector |Ψ on one of these basis, we obtain its projection onto this plane. Taking this vector and renormalizing gives us the desired vector |𝝍 . The vector |𝜓 can be decomposed on both orthogonal basis. |𝜓 = α|𝑢𝐴 + α⊥ |𝑢𝐴⊥ = β|𝑢𝐵 + β⊥ |𝑢𝐵⊥ The coefficients α, α⊥ , β and β⊥ are obtained by projecting the state vector |Ψ on both basis vectors and then normalizing to unity. QUANTUM MODEL FOR HAL : QUERY OPERATORS We want to define Query operators. The query operators 𝐴 and 𝐵 are defined +1 state that corresponds to the word meaning we are interested. −1 in the orthogonal direction. When applied to the document state vector |𝜓 defined before : 𝐴|𝜓 = α|𝑢𝐴 − α⊥ |𝑢𝐴⊥ 𝐵 |𝜓 = β|𝑢𝐵 − β⊥ |𝑢𝐵⊥ This action is analogous to the spin Pauli matrix 𝜎𝑧 , and we can associate it to 𝐴 in the basis {|𝑢𝐴 , |𝑢𝐴⊥ } 𝐴 = 𝜎𝑧 = 1 0 0 −1 Other query operators can be defined. We choose 𝐴𝑥 and 𝐵𝑥 . The action of 𝐴𝑥 on |𝜓 gives on the {|𝑢𝐴 , |𝑢𝐴⊥ } basis: 𝐴𝑥 α|𝑢𝐴 + α⊥ |𝑢𝐴⊥ = α⊥ |𝑢𝐴 + α|𝑢𝐴⊥ This action corresponds to switching the components and is equivalent to the spin Pauli matrix 𝜎𝑥 𝐴𝑥 = 𝜎𝑥 = 0 1 1 0 QUANTUM MODEL FOR HAL : QUERY OPERATOR BASIS REPRESENTATION We choose the basis associated to word 𝐴 , {|𝑢𝐴 , |𝑢𝐴⊥ } , and write the operators with respect to this basis. Using the transformation matrix 𝑀 from the 𝐴 basis to the 𝐵 basis 𝑀= 𝑢𝐵 𝑢𝐴 𝑢𝐵⊥ 𝑢𝐴 𝑢𝐵 𝑢𝐴⊥ 𝑢𝐵⊥ 𝑢𝐴⊥ = 𝑝 − 1 − 𝑝2 1 − 𝑝2 𝑝 where 𝑝 = 𝑢𝐵 𝑢𝐴 is the scalar product here a positive number smaller than 1 We obtain the 𝐵 matrix form in the {|𝑢𝐴 , |𝑢𝐴⊥ } basis associated : 𝐵= 2𝑝2 − 1 2𝑝 1 − 𝑝2 2𝑝 1 − 𝑝2 1 − 2𝑝2 BELL PARAMETER CALCULATION USING QUERY OPERATORS Bell tests are usually a proof of a non-separability of the combination of two different systems. We define a parameter 𝑆𝑞𝑢𝑒𝑟𝑦 that can be understood as the sense associated to a word A in a document in correlation with the sense of another word B. The Bell parameter 𝑆𝑞𝑢𝑒𝑟𝑦 is the combination quantum mean values with different query operators which can be considered as measuring devices: 𝑆𝑞𝑢𝑒𝑟𝑦 = 𝐴𝐵+ + 𝐴𝑥 𝐵+ + 𝐴𝐵− − 𝐴𝑥 𝐵− Using specific operators associated to words A and B . 𝐴 ; 𝐴𝑥 ; 𝐵+ = − 1 2 𝐵 + 𝐵𝑥 ; 𝐵− = 1 2 𝐵 − 𝐵𝑥 This particular operator choice is inspired from the usual example that maximizes the violation of the Bell inequalities . We calculate quantum mechanical mean values over vector |𝜓 using the Born rule. For example: 𝐴𝐵 𝜓 = 𝜓 𝐴𝐵 𝜓 BELL PARAMETER CALCULATION: Q-HAL ALGORITHM Input Document Construction of a “clean” (no punctuation marks) sequence of words, including eventual repeated words: Doc list. Construction of the “Dictionary”: sequence of non repeated words: Dic list. Window size l Construction of a primitive HAL matrix: for each word of the Doc list a window of length l is associated and all the scores of the words within it are collected in a matrix. The entry for each score is determined by the position of the words in the Dic list. Complete HAL matrix is obtained by summing this matrix with its transpose. Normalization of each row vector. Determination of the state of the system by summing over all vectors and normalizing. Calculation of the expected values of the defined operators and the Bell parameter. New window size l+1 Plot Flow diagram of the Quantum HAL algorithm. The algorithm was implemented using Python programming language along with the string module and pylab. Our approach presented here can be perceived as an experiment done on objects outside the domain of physics. QUERY 1 : WORD “REAGAN” IN THE CONTEXT OF WORD “IRAN”. The Bell parameter function of the window size starts from zero and increases until it reaches a maximum (the Tsirelson's bound 2 2) then drops again. (for 3 documents) The document “Iran” always gives a constant value of 2. Here one of the words (“Reagan”) is missing. This suggests that each document has an optimal HAL window size that maximizes the parameter 𝑆𝑞𝑢𝑒𝑟𝑦 . The “sooner” a peak appears the less interaction, in the sense of window length, is needed to get higher correlation between the two words. Bearing this in mind, the document “IranContra affair” is clearly the one selected by the model. QUERY 2: TEST ON THE POLYSEMY OF THE WORD “ORANGE” Test on the polysemy of the word “orange” and associated concepts. In this example we are interested in the ambiguity between the meanings color and fruit. We also associate the concept of juice. The query “Orange - Fruit” presents the first peak around 𝑙 = 22 for the document “Orange color”, the second in 𝑙 = 29 for the document “Orange fruit”, then for 𝑙 = 40 the document “Orange juice” and very far away the document “Juice”. Regarding the peaks we find the expected order for the documents: “Orange juice”, “Orange Fruit”, “Juice” and “Orange Color”. QUERY: PATHOLOGICAL TEXT EXAMPLE Queries on words A1 and A100 in a text of 5000 words, for text repetition periodicities of 100, 150 and 200. We made 2 word queries on « pathological » documents consisting in texts with repeating periodic structure based on the same original document. The curves still peak at the Tsirelson’s bound and also present other effects probably due to the repetition period. COMMENTS ON RESULTS The results show Bell parameter that peaks up to the maximal value of Sbell = 2√2, (the Tsirelson’s bound). We found that the Bell parameter is strongly dependent on the HAL window size. There is an optimal window size that maximizes Sbell. Reminiscent of what was already noticed (Bruza) a possible explanation : if the window size is set too large, spurious co-occurrence associations are represented in the matrix if the window size is too small, relevant associations may be missed. Comparing different documents, the one with the first appearing peak seems to be the more relevant. SOME CONCLUSIONS AND PERSPECTIVES The main feature in relation to Quantum Theory explored in this work is the violation of the Bell inequalities which can be related to entanglement and nonlocality. The results show always correlation on two words due to Bell inequality violation up to the maximal value of 𝑆𝐵𝑒𝑙𝑙 = 2 2, (the Tsirelson’s bound). We introduced a new tool which has connections with the Quantum Theory: Query Operators. It is not clear how to interpret the Bell inequality violation here and what is the meaning of the optimal length that maximizes the Bell parameter. HAL constitutes a good « playground » for doing Quantum-like experiments. We believe that it should be possible, after much experimentation on different documents, to introduce new families of query observables adapted to different purposes and contexts in Information Retrieval. Can entanglement give a measure of query relevance? Thank You